Effective Data Driven Coding Unit Size Decision Approaches for HEVC INTRA Coding

Quality of Experience (QoE)-Aware Fast Coding Unit Size Selection for HEVC Intra-Prediction

“…Following a similar approach, Shen and Yu [14] propose an early termination algorithm for recursive CU evaluation using Support Vector Machines (SVMs). In addition, Zhang et al [10] propose an early termination/early split mechanism where multiple SVMs are utilised. The two-stage approach proposed in [10] utilises offline trained models and weight parameters for the decision making process.…”

“…In addition, Zhang et al [10] propose an early termination/early split mechanism where multiple SVMs are utilised. The two-stage approach proposed in [10] utilises offline trained models and weight parameters for the decision making process. However, the offline computation of SVM weights and hyperplane and support vectors make the CU split decisions obtained through the models less content-adaptive.…”

“…The split likelihood of a CU can be modelled using probabilistic models [7,35,36] as well as machine learning methods. In the case of the latter, recent literature typically uses supervised learning algorithms such as decision trees [17,38], neural networks [34], logistic regression, and SVMs [10,14,39]. However, integrating an inference engine of machine learning models into the encoding loop such that the time taken for decision making is minimal, is crucial a for fast CU size selection algorithms (such as proposed in this work).…”

“…where, α i , β i , and γ i correspond to the texture complexity, RD cost pre-analysis of current CU, and context information, respectively, for the ith CU depth level [10]. In this case, the texture complexity (α i ) is calculated using the mean absolute difference (α 1 ) of the current CU and the texture difference between the current CU and its four sub CUs (α 2 ).…”

“…In this context, the recent literature is composed of numerous statistical inference-based methods [6,9], texture properties [10,11], and machine learning-based algorithms [12][13][14] that focus on reducing the encoding complexity of HEVC intra-predicted frames. However, rigid decision trees and pre-trained static prediction models make these algorithms less adaptable to the dynamic changes of the video content characteristics.…”